"Towards VLSI Spiking Neuron Assemblies as General-Purpose Processors"

02. April 2012

Neuromorphic engineering is a multidisciplinary approach aiming to design novel electronic devices by mimicking the computationalprinciples observed in neurobiology. One of its subfields is dedicated to designing multi-neuron chips with the aim of providing the computational abilities of biological information processing systems.
Although the technical implementations of such devices are rapidly improving, a computational methodology able to achieve desired complex cognitive functionalities is still lacking.
Here, we introduce a model-based procedure which is able to configure a real-time neuromorphic multi-chip system to generate the behavioral states required for solving target state-dependent tasks. The neural model expresses behaviors analogous to those programmed on classical finite state machines. Its architecture consists of recurrently connected neural networks using attractor dynamics to maintain persistent activity states, and to drive transitions between them.
The resulting system is able to interact with the environment via event-based sensors and attentional pre-processing to robustly carry out a wide range of complex sensorimotor tasks.